1. Field of the Invention
The invention relates to noise discrimination in signal detection and processing.
2. Description of the Related Art
FIG. 1 is a block diagram of a conventional real-time frequency domain signal processing system 10 employing what is sometimes referred to as the frequency sub-band method or the frame-overlap-and-add method. This method uses a circuit 11 to divide incoming sampled temporal signal information into blocks of data referred to as frames. The sampled data can be provided directly from a digital sensor or other processing system, or can be provided from an analog sensor or processing system via a standard Analog-to-Digital conversion (A/D or ADC) method (not shown). The frames can be adjacent or overlapping. Since the data are samples of time domain data, all samples within a frame have no imaginary component, and the data is strictly “real.” If required by the application, these frames of data then may be multiplied in a multiplication circuit 12 by an analysis window 14a to reduce artifacts that can be introduced by subsequent transformation of the sampled time data into the frequency domain. Subsequently, the windowed frames are transformed to the frequency domain by any one of the many such transformations known to those of ordinary skill in the art, such as for example the Hartley transform, the Wavelet transform, or the like. The most commonly used of these transformations is the Fourier transform. Since the data is sampled and digitized, the DFT, or Discrete Fourier Transform, is used in these cases, with a preference for using one of the fast-to-compute versions of this transform, known as the Fast Fourier Transform or FFT, represented at circuit 16.
Although there are choices for the analysis window, such as the Hanning window, that will reconstruct the time domain signal accurately without the added complexity and computational cost of a synthesis window, such analysis windows suffer from accuracy compromises to achieve the improved efficiency. Generally, a separate synthesis window 14b is applied by multiplication before the signal is reconstructed by the overlap and add circuit, 19 (as shown in FIG. 1) to overcome these compromises, but at added cost.
Once in the frequency domain, the data is represented by complex numbers containing both a “real” and an “imaginary” component. These complex numbers, one for each frequency “bin” of the transform, represent the magnitude and relative phase angle of the temporal input signal data averaged over the time interval contained within the length of the frame (and weighted by the windowing function) as well as over the range of frequencies contained within the bandwidth of the “bin.” It is this input transform data that is then processed at circuit 17 by a selected process to create an output transform of processed frequency domain data.
Once the data is processed, the standard frequency domain method then calls for inverse transformation of each frame of processed data to create a string of processed time domain frames of “real” data. Circuit 18, emoting an inverse fast Fourier transform (IFFT) process, performs this objective. If a synthesis window 14b is used, then it is applied at circuit 13 by multiplication of the output frame of time domain data with the selected synthesis window: otherwise the output frame of data from circuit 18 is passed directly to circuit 19. Alternatively, the frequency domain representation of the synthesis window can be applied to the output from the signal process 17 by convolving the output from the process with the transformed synthesis window before performing the inverse Fourier transform at circuit 18. The time domain frames are subsequently re-assembled by circuit 19 by performing concatenating or overlapping-and-adding of the frames of processed real-time data to create the final digitized and sampled temporal output signal waveform containing the processed signal information. Of course, this sampled signal can be, and often is, converted into an analog signal by the use of a standard Digital-to-Analog conversion (D/A or DAC) method (not shown) so that the processed output signal can be used in myriad applications, such as scientific measurement, telephony, entertainment systems, communication systems, and so on.
Alternatively, the process can be applied in the time domain, wherein, for example, the input signal, either analog or digitized, is passed through a bank of bandpass frequency discrimination filters (either analog or digital as appropriate). The outputs of each of the frequency filters is subsequently processed, and the processed signals are then combined to form a processed output signal by adding the processed signals together.
FIG. 2(a) shows the elements of a conventional prior art beamforming system, where a sensor system 21 provides two or more input signals 22 that are time-aligned for the signal of interest. For best performance, these sensor signals should have matched sensitivity for all signals. The input sensor signals 22 provide the input data for the vector summing beamforming process of the system, as shown at circuit 23.
Although the vector summing process 23 is often performed as a vector average, a vector average is simply a vector sum divided by a scalar number, and will simply be referred to hereinafter as a vector sum.
Consider one of the simplest beamforming sensor systems, the two-element broadside array 30 shown in FIG. 3. The two sensor elements 32 and 34 of this array are located on the axis X It is well known that such a beamforming system can be steered using conventional signal delay methods. In particular, conventional beam steering is accomplished by varying relative phases of the input signals in such a way that the incoming signal pattern is reinforced in a desired direction and suppressed in undesired directions. The phase change is equivalent to a time delay—that is, the phase change at each frequency is a fixed offset, and the phase change over frequency is linear. However, for simplicity here it is assumed that the signal source of interest lies on the sensitivity axis I of the array—that is, that the two sensor signals are appropriately time delayed so as to be time-aligned for the desired signal of interest. When the sensor elements 32 and 34 are omni-directional and spaced one-half wavelength apart (180 electrical degrees), the two-element broadside beamforming system, as shown in FIG. 2(a), outputs a signal that is directly proportional to the vector sum of the two sensor element signals. This output has a sensitivity beam pattern resembling a figure-eight—that is, one having two sensitivity lobes 35 and 36 as shown in FIG. 3. These lobes are maximum in the on-axis direction, but are zero at ±90°. azimuth directions (in the directions of axis X). These are the directions at which the electrical phase difference between the sensor's signals is ±180° and therefore where the signals cancel when summed together. The resulting low sensitivity regions 37 and 38 are referred to as “nulls.”
To improve the directionality of a sensor system normally implies narrowing the width of the main lobe(s) of sensitivity, which in FIG. 3 is either lobe 35 or 36 (or both). In a conventional beamforming system, narrowing of the main sensitivity lobe is accomplished by incorporating additional sensor elements to enlarge the array, thereby increasing the acceptance aperture that concomitantly reduces the beam width. However, there are costs to this approach, including the additional sensor elements and associated amplifiers and A/D converters (in a digital system) or filters (in an analog system), the added computational costs for processing all the sensor signals, the result that the beam pattern becomes complex with many added side lobes in which the sensitivity of the system to unwanted signal sources is relatively high (that is, the system has relatively low noise immunity), the large physical size of the sensor array, and non-uniform frequency response for off-axis signals, among others.
For these reasons, another method called “super resolution” beamforming has been employed, wherein the increased aperture is filled with additional sensor elements, but the elements are non-uniformly spaced and the resulting sensor signals are non-uniformly weighted in amplitude. In such a system (not shown), the width of the main lobe of sensitivity can be more greatly narrowed as compared to a similar beamforming system with uniformly spaced sensor elements. However, to be successful the super resolution approach still requires a great number of sensor elements and associated circuitry and suffers from significantly increased computational costs, high side lobe sensitivity, large physical size, and non-uniform off-axis frequency response.
In order to address the side lobe pickup problem, another method has been employed in which additional beamformer systems are used with the same set of array sensor signals. The additional beamformers create sensitivity beams that are in the directions of the side lobes of the main beamformer. The output signals from these additional beamformers are then scaled and subtracted from the output signal from the main beamformer in order to partially cancel the main beam former's side lobes. In general, although the side lobes can be reduced with such an approach, the tradeoffs include a wider main lobe, high complexity and cost, and the retention of a high number of sensors.
Yet another category of conventional beamformer is the generalized side lobe canceller (GSC) where a multiple sensor system is combined with a null-steering method. In this technology, the sensitivity toward the desired source is maintained constant while one or more of the nulls are steered toward detected off-axis noise sources. Examples of this type of beamforming system are the well known Griffiths-Jim beamformer and the Frost beamformer. In this type of beamforming system the number of discrete noise sources that can be nulled is equal to the number of independently steerable nulls, and the number of independently steerable nulls is equal to one less than the number of sensors. Thus, to be effective in most real-life situations where there are numerous noise sources and multiple-reflections of those noise sources, the number of sensors must be large, along with the associated high system complexity, large compute power requirement, and high cost. Further, such systems, because the nulls are very narrow, require adaptive circuit techniques to accurately center the nulls on the noise source directions, and these adaptive methods are slow to adapt, allowing significant noise to pass during the adaptation time.
One common characteristic of these prior art systems is that the null or nulls created by these methods are quite narrow. As more sensor elements are incorporated, more nulls are created and the numerous resulting nulls are narrower yet.